Segmentation of Textured Image Based on Gauss-MRF in Overcomplete Brushlet Domain

The focus of this paper is the improvement of the quality of texture image segmentation. We proposed a new unsupervised segmentation method based on Overcomplete Brushlet transform and Gaussian Markov Random Field. A texture image was transformed to Overcomplete Brushlet domain, in order to extract its high dimensional singularity information. In view of the influences of the spectral information and the spatial correlations between pixels to the segmentation result, Markov Random Filed model is used in the process of both feature extraction and region segmentation: Gauss Markov model is used to evaluate the arguments of the feature field; the probability of the marker field is calculated through Gibbs distribution function based on the second order neighborhood system of MRF. MAP criterion is adopted to obtain segmentation results. We did a lot of contrast experiments, using this paper's algorithm, Markov Random Field algorithm in wavelet domain and Markov Random Field algorithm in Brushlet domain. Those experiment results indicate that this paper's algorithm is an effective segmentation algorithm for it can detect better texture direction information and keep better regional consistency than other two traditional algorithms.